Batch structured AI calls via OpenRouter, Data Tables, and webhooks

Quick overview

Give an AI coding agent one authenticated webhook for structured LLM batches. The agent prepares prompts and JSON schemas in a script; n8n processes the first 50 calls in parallel, returns ordered results, and records who used the service.

How it works

1. The Webhook accepts a POST request with an X-API-Key header and a calls array. Each call contains systemPrompt, userMessage, a JSON-encoded jsonSchema string, and an aiMode of small or big. Use one mode per request.
2. Check Auth* looks up the token in the users Data Table. Unknown tokens receive HTTP 401. For accepted requests, Insert log stores the user's email and the full submitted calls payload, including prompts, schemas, and modes.
3. Prepare Calls* restores the request body after authentication, and Split Calls creates one item per call. Limit to 50 passes the first 50 items. It silently drops any extras, so clients should split larger jobs and check result counts.
4. Basic LLM Chain* processes the retained items in parallel. Model Selector routes small to GPT 6 Luna and big to Gemini 3.8 Flash. Structured Output Parser applies each call's schema. Adjust the chain's batch size and delay to match provider rate limits.
5. Aggregate Responses* collects the results. Respond to Webhook returns ordered {query, output} entries, one per processed call. query echoes the submitted input; callers should inspect each output and handle item-level failures.

Setup

1. Create a users Data Table with string columns email and auth_token. Create a logs Data Table with string columns user and log in the same n8n project.
2. Add one users row per person. For this example, generate a unique 12-character alphanumeric auth_token with a cryptographically secure random generator and deliver it privately. Your organization can use its own token system instead.
3. In Check Auth, select the imported workflow itself. In Get team member, select the users table and filter on auth_token using ai_auth_key. In Insert log, select the logs table.
4. Configure an OpenRouter credential in GPT 6 Luna and 3.8 Flash. If you replace either model or provider, reconnect the model nodes and verify the small and big rules in Model Selector.
5. Publish the workflow and copy the production URL from Webhook. Give each user that URL and their personal token through a verified private channel. They store them as unquoted EXTERNALIZED_AI_URL=... and EXTERNALIZED_AI_API_TOKEN=... values in a local .env.
6. Run one cheap test call, then a small batch. Check the JSON response, the 401 response for a bad token, and the logs table. Logs retain submitted prompts and source text; restrict table access and set an appropriate retention policy.

Requirements

• An n8n instance with Data Tables, LangChain nodes, and a reachable production webhook.
• An OpenRouter account and credential, or replacement model credentials configured in the model nodes.
• A secure way to issue and privately deliver a personal token to each user.

Customization

• Replace GPT 6 Luna and Gemini 3.8 Flash with other models or providers while keeping the small and big contract for callers.
• For per-call model selection, measured cost, reasoning effort, or other model types, replace the model branch with direct OpenRouter API calls. This is an extension to the two-model template.
• Replace the users table with your own identity store. A separate onboarding workflow could deliver tokens through a private Slack DM after verifying the requester's identity.

Additional info

jsonSchema is a JSON-encoded string, not a nested object. Send one aiMode per request. Limit to 50 silently drops extra calls, and individual calls may fail without a usable output. The response echoes inputs under query. The logs table stores all submitted calls, including those beyond the 50 processed items, with prompts and source text. Restrict access and retention. Model outputs and measured cost are not in the log. OpenRouter docs: https://openrouter.ai/docs/quickstart.